nvcc outputs mangled names CUDA compilation

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi All,

I have some C++ source files which call external CUDA functions. But when I try linking the cpp obj files with cu obj files g++ linker throws lot of undefined symbols errors. Did a nm on the .cu obj files and it returns some thing like __Z22Pf where the g++ linker is searching for _.
Any ideas on how to disable name mangling when compiling with nvcc.

Thanks
Radha

2

Wrap the functions like so:

extern "C"

{

int someFunct(...)

{

...

}

}

or, simply do:

extern "C" int someFunct(..)

{

..

}

No more name mangling.

3

Thanks for the reply. I did wrap all the cuda functions calls from cpp files as extern “C”. Do you mean to say wrap functions declaration in .cu files as extern C?

Regards

Radha

4

alternately, just don’t use extern “C” anywhere and stick to extern if you’re linking with C++.

5
6

Tried that didn’t seem to work.

7

Thanks for the ideas. I did wrap all the cuda functions in .cu files and the cuda function calls in .cpp files as extern “C” but still g++ linking complains about undefined symbols. And nm on object files shows the function names are not mangled. Any ideas please share.

Thanks

Radha

8

Got it…extern “C” worked but in addition to that while compiling with nvcc added -m32 and while linking -m32 and -arch i386 (Mac OS Snow Leopard). I was linking with OpenCV libs as well and had to re-compile OpenCV libs with -m32 and -arch i386 in order to link properly. Previously when linking to OpenCV libs compiled without -m32 and -arch i386 the linker could not see any of the OpenCV symbols.

Thanks